The reliability of multiple regression and an alternative method for extracting task-specific exposure estimates from time-weighted average data.

The reliability of multiple regression analysis as a method for determining task-specific exposures from multi-task time-weighted average data was evaluated in comparison with the alternative P-screen method. The performances of the two methods were tested using simulated sample data that were calculated as averages over six tasks, where task-specific concentrations drawn randomly from lognormal distributions were weighted by randomly generated task time-weights. Data sets consisted of 20 or 100 simulated samples. The simulated data sets conformed to requirements inherent in the P-screen method that at least one task be absent from each sample and each task be absent from at least one sample. In thousands of Monte Carlo trials under various conditions, the two methods were found to perform equally well when dichotomous task measures (occurrence/ nonoccurrence) were used. Combining the two methods did not improve reliability appreciably, suggesting that the methods are effectively equivalent when dichotomous task measures are used. When task durations were used as the regressors or time-weights, multiple regression was found to be more reliable than P-screen. It is well recognized that incidental or fundamental collinearities between regressors may undermine multiple regression analyses. The P-screen-related restrictions on the task structure of data sets reduces the potential for problems arising from such collinearities. However, the use of multivariate analysis of multiple-task samples will always be an imperfect substitute for single-task sampling.